Fig. 7: Schematic representation of the two-layered variational deep Gaussian process (DGP) architecture.

Each variational multi-task Gaussian process (MTGP) layer reduces dimensionality from input to latent representation, governed by the reduction parameter.